[Bug target/59788] Mixing libc and libgcc_s unwinders on 64-bit Solaris 10+/x86 breaks EH

[ro at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59788

--- Comment #2 from Rainer Orth <ro at gcc dot gnu.org> ---
Created attachment 31843
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=31843&action=edit
initial patch

The attached initial patch works fine and fixes the testcase from the PR, but
is
unacceptable as is.

Initially, I attempted to force direct binding to the libgcc_s unwinder by
linking
with a special mapfile if linking with -lgcc_s.  This doesn't work since
libtool
builds shared libraries with -shared -nostdlib, adding -lgcc_s itself.

The second attempt (attached) instead uses that mapfile when -shared is passed.
When libtool adds -lgcc_s -lc -lgcc_s itself, it usually `optimizes' this into
-lc -lgcc_s.  When direct binding is in effect due to the mapfile, we achieve 
exactly what this patch intends to avoid, namely binding to the partial
unwinder
in amd64 libc.  This can be avoided by patching libtool to disable this
optimization
on *solaris2*, which this patch does.

While this works fine for the gcc runtime libraries, as can be observed with 
elfdump -y, gcc 4.9 cannot be released with this patch included: even if the
libtool (ltmain.sh actually) patch were accepted upstream, every package built
with libtool will contain an old copy without that change, causing the breakage
this patch is designed to avoid.  While the optimization can be avoided by
invoking libtool with --preserve-dup-deps, this seem unacceptable, as would
requiring users to run libtoolize from a hypothetical new libtool release with
the ltmain.sh patch.